ChatMemory is persisted over request boundaries

[andreas-eberle]

I was not looking a bit more into the examples from the guides page and especially the chat memory management (https://docs.quarkiverse.io/quarkus-langchain4j/dev/ai-services.html#memory). Unfortunately, I got a bit confused when I tested it as it seems to behave quite different from what I initially expected. Hopefully you can help to clarify things.

In the example code, the ChatMemoryBean is created as a RequestScoped bean. However, when you just use it, its get method is only called once with default as memoryId. At the end of the first request, it's close method is also called (as expected) and the memories map is freed. However, when you do the second request, the bean is not used any more and no new memory is created. Instead, all other requests just use the same memory.

Is this expected? If yes, why is the bean even request scoped?

One thing I noticed is that the get method is called by dev.langchain4j.service.AiServiceContext. And AiServiceContext itself has a map doing basically the same as ChatMemoryBean in the sense that it has a map from memory id to ChatMemory. I think this is where the memory lives longer than the request scope.

see also https://quarkusio.zulipchat.com/#narrow/stream/187030-users/topic/Quarkus.20Langchain.3A.20Chat.20Memory.20Management.20Questions

[geoand]

This is definitely a bug. I think I have a fix but it will have to wait for #83 to go in first

[andreas-eberle]

I think this might also play into #47.

[geoand]

Yeah, that's a more general thing we need to address.

[geoand]

#96 should take care of it